* fix(runinfra): declare image input on Qwen3.8 27B and the accepted "high" effort synonym
Two facts in the RunInfra entries stopped being true after they were written.
Qwen3.8 27B: image input went live on this host on 2026-08-20, so declaring
text-only understated the model. modalities.input is still overridden rather
than deleted because the host accepts text and image and does not serve video,
which the base model declares; `attachment` is dropped because it now matches
the base model. Confirmed on our own published provider feed, which derives
the accepted set from the same registry the gateway enforces:
https://runinfra.ai/api/providers/openrouter/models reports text+image.
Qwen3.8 27B and Qwen3.8 2.4T A95B: the gateway now accepts reasoning_effort
"high" and folds it to "xhigh" before dispatch, so the 27B's note that "high"
returns HTTP 400 is stale. Neither publishes it as a separate rung, following
the convention already used on DeepSeek V4 Flash: a value that folds onto
another is documented in the comment, not listed as a distinct behaviour.
"high" is only the OpenAI ladder's word for the maximum and these serves top
out at "xhigh". "max" is still rejected on the 27B and is still not published,
and "none" is still rejected on the 2.4T, which is deliberate: a caller asking
to switch reasoning off should learn the model has no off switch rather than
be silently served the most expensive setting.
No price, limit or capability boolean changes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(runinfra): correct two dates that named the wrong day
Both claims in this PR were true; both carried a date that git contradicts.
The 27B's image input shipped 2026-08-18 (58591f8b9), not 2026-08-20.
2026-08-20 is Ornith's vision date, a different model.
The 2.4T's acceptance of the "high" effort synonym shipped 2026-08-21
(af3d1b356), not 2026-08-23. 2026-08-23 is the 27B's own fold (4046a8e00),
which line 12 of the 27B file already dates correctly; two model changes had
been collapsed onto one date.
Nothing published changes: no price, no limit, no capability boolean, no
reasoning value. Only the evidence comments, which are the part of these
entries a reviewer is asked to trust.
This provider's whole argument to these maintainers is dated, verifiable
measurement. A comment that misdates its own evidence undermines that more
than the field it decorates is worth.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Jaber Jaber
9647698a74